Inventing the Future of Computing
Parallella: A $99 Open Hardware Parallel Computing Platform Andreas Olofsson [email protected]
IPDPS May 22th, Cambridge, MA Adapteva Achieves 3 “World Firsts” 1. First commercial processor to reach 50 GFLOPS/W
2. First mobile processor with an open source OpenCLTM SDK 3. First semiconductor company to successfully crowd‐source project
Copyright © Adapteva. All rights reserved. 2 Adapteva’s Goals in 2008
• A C/C++ programmable multicore processor
• Scalable to 1000’s of cores on a chip
• Native IEEE floating point support
• Easy to Use
• 50 GFLOPS/Watt in 65nm
Copyright © Adapteva. All rights reserved. 3 Our Inspiration
Teraflop Intel Tile (2007) Tilera (2006) RAW MIT Transputer (1997) http://www.adapteva.com/white‐papers/the‐ Inmos siren‐song‐of‐parallel‐computing/ (1984) Copyright © Adapteva. All rights reserved. 4 Our guiding light
Parallel Efficient
Hetero‐ Robust geneous
Copyright © Adapteva. All rights reserved. 5 Any Reason to Think the Future of Computing is NOT Parallel?
No Computing Parallel Computing
No Electronic “Von Neumann Age” Parallel Computing Computing Serial Computing 2013‐?? ‐1943 1943‐2013?
Copyright © Adapteva. All rights reserved. 6 A Practical Start: True Heterogeneous Computing O/S SYSTEM‐ON‐CHIP Application BIG BIG BIG BIG “TheCPU CPU CPU CPU Graphics Joker” Weird Math Math 1000’s of FPGA GPU Analog small RISC CPUs
Copyright © Adapteva. All rights reserved. 7 The Epiphany Coprocessor 1.6 GFLOPS 32‐128KB Per Core @ Local ~25mW Memory
Packet Based Network‐On‐Chip With 100GB/s Bisection BW
Coprocessor for MIMD/Task-Parallel <20pJ / FLOP ! ARM/x86 Host Accelerator
Copyright © Adapteva. All rights reserved. 8 Pragmatic Architecture Tradeoffs IN OUT • Dual issue RISC processors • Any special purpose instructions • 64 entry register file • Hardware caching • Shared memory architecture • SIMD • 32-128KB per core memory • Optimized remote read accesses • Multi-banked local memory • Memory management unit • Packet based Mesh NOC • Strict memory order model • 32 Bit IEEE float/int arithmetic • Memory protection • Timers, Interrupts, DMAs
Copyright © Adapteva. All rights reserved. 9 How the $#@% Do We Program This Thing?
Copyright © Adapteva. All rights reserved. 10 The Current State of Parallel Programming
Source: Github
How To Make Every Programmer a Parallel Programmer?
Copyright © Adapteva. All rights reserved. 11 Parallel Programming Frameworks
Erlang SystemC Intel TBB Co‐Fortran Lisp Janus Scala Haskell Pragmas Fortress Hadoop Linda Smalltalk CUDA Clojure UPC PVM Alef Julia OpenCL Go X10 Posix XC Occam OpenHMPP ParaSail APL Simulink Charm++ Occam‐pi OpenMP Ada Labview Ptolemy StreamIt Verilog OpenACC C++ Amp Rust Sisal Star‐P VHDL Cilk Chapel MPI MCAPI ?????????
Copyright © Adapteva. All rights reserved. 12 Stupid Hurdles That Get in the Way of Progress • Proprietary SDKs and programming frameworks
• Lack of datasheets/documents
• Closed source drivers
• Expensive lock-in hardware
• NDA requirements
• Exlcusive access
Copyright © Adapteva. All rights reserved. 13 The Parallella Project Guidelines
• A $99 single board ”parallel” computer that runs Linux
• Open source (SDK, board files, drivers) (github.com/parallella)
• Open documentation (adapteva.com/all-documents)
• Open to all (forums.parallella.org)
Copyright © Adapteva. All rights reserved. 14 The Parallella Board Zynq dual core ARM- A9 16-core (with FPGA Logic) Epiphany Coprocessor Gigabit Ethernet uUSB
uHDMI
uUSB uSD 5V DC
1GB SDRAM Copyright © Adapteva. All rights reserved. 15 The Parallella Backside (optional) Epiphany North FPGA Logic Connector 6 GB/s ! BW Connector
Instrumentation Connector Epiphany South Copyright © Adapteva. All rights reserved. Connector 16 Parallella Kickstarter Campaign
• 5,000 customers • 6,300 boards ”pre-sold” in 4 weeks • 67 countries, all 50 US states • 50-75% of backers are developers • 11,000 more signups since Jan 1st
• Backer Application Interest: • Software Defined Radio • Cryptography • Ray tracing/rendering • Parallel computing research • Image processing • Distributed Computing • Robotics • Machine Learning • Gaming • HPC
Copyright © Adapteva. All rights reserved. 17 The Parallella 1% Academic Program P P P P P P P P P P • Adapteva will donate (at least) 1 board for P P P P P P P P P P every 100 boards sold! • Open to all academic institutions active in P P P P P P P P P P parallel computing research & education P P P P P P P P P P
P P P P P P P P P P
P P P P P P P P P P
P P P P P P P P P P
P P P P P P P P P P P P P P P P P P P P Program starts June 1st, 2013! P P P P P P P P P P
Copyright © Adapteva. All rights reserved. 18